The document discusses advancements in Apache Spark 2.x, particularly focusing on structured streaming and deep learning capabilities. It highlights the introduction of a high-level API for streaming that simplifies processes and allows for end-to-end applications, as well as the integration of deep learning frameworks to enhance usability. Key features include improved performance, cost efficiency, and the goal of making deep learning accessible to a broader range of users.

1. Deep Learning and
Streaming in Apache Spark
2.x
Matei Zaharia
@matei_zaharia

2. Welcome to Spark Summit Europe
Our largest European summit yet
102talks
1200attendees
11tracks

3. What’s New in Spark?
Cost-based optimizer (Spark 2.2)
Python and R improvements
• PyPI & CRAN packages (Spark 2.2)
• Python ML plugins (Spark 2.3)
• Vectorized Pandas UDFs (Spark 2.3)
Kubernetes support (targeting 2.3)
0
10
20
30
40
50
Time(s)
Spark 2.2
Vectorized UDFs
0
25
50
75
100
125
Q1 Q2
Time(s)
Pandas
Spark

4. Spark: The Definitive Guide
To be released this winter
Free preview chapters and
code on Databricks website:
dbricks.co/spark-guide

5. Two Fast-Growing Workloads
Both are important but complex with current tools
We think we can simplify both with Apache Spark!
Streaming Deep
Learning
&

6. Why are Streaming and DL
Hard?Similar to early big data tools!
Tremendous potential, but very hard to use at first:
• Low-level APIs (MapReduce)
• Separate systems for each task (SQL, ETL, ML,
etc)

7. Spark’s Approach
1) Composable, high-level APIs
• Build apps from components
2) Unified engine
• Run complete, end-to-end apps
SQLStreaming ML Graph
…

8. Expanding Spark to New
Areas
Structured Streaming
Deep Learning
1
2

9. Structured Streaming
Streaming today requires separate APIs & systems
Structured Streaming is a high-level, end-to-end API
• Simple interface: run any DataFrame or SQL code incrementally
• Complete apps: combine with batch & interactive queries
• End-to-end reliability: exactly-once processing
Became GA in Apache Spark 2.2

10. Structured Streaming Use
Cases
Monitor quality of live video streaming
Anomaly detection on millions of WiFi hotspots
100s of customer apps in production on Databricks
Largest apps process tens of trillions of records per month
Real-time game analytics at scale

11. KTable<String, String> kCampaigns = builder.table("campaigns", "cmp-state");
KTable<String, CampaignAd> deserCampaigns = kCampaigns.mapValues((value) -> {
Map<String, String> campMap = Json.parser.readValue(value);
return new CampaignAd(campMap.get("ad_id"), campMap.get("campaign_id"));
});
KStream<String, String> joined =
filtered.join(deserCampaigns, (value1, value2) -> {
return value2.campaign_id;
},
Serdes.String(), Serdes.serdeFrom(new ProjectedEventSerializer(),
new ProjectedEventDeserializer()));
KStream<String, ProjectedEvent> filtered = kEvents.filter((key, value) -> {
return value.event_type.equals("view");
}).mapValues((value) -> {
return new ProjectedEvent(value.ad_id, value.event_time);
});
KStream<String, String> keyedData = joined.selectKey((key, value) -> value);
KTable<Windowed<String>, Long> counts = keyedData.groupByKey()
.count(TimeWindows.of(10000), "time-windows");
streams
Example:
Benchmark DataFrames
events
.where("event_type = 'view'")
.join(table("campaigns"), "ad_id")
.groupBy(
window('event_time, "10 seconds"),
'campaign_id)
.count()
Batch Plan Incremental Plan
Scan Files
Aggregate
Write to Sink
Scan New
Files
Stateful Agg.
Update Sink
automatic
transformation

12. 4xlower cost
Structured Streaming
reuses the Spark SQL
Optimizer and Tungsten
Engine.
https://data-artisans.com/blog/extending-the-yahoo-streaming-benchmark
Performance:
Benchmark System Throughput
700K
15M
65M
0
10
20
30
40
50
60
70
Kafka
Streams
Apache Flink Structured
Streaming
Millionsofrecords/s
4xfewer nodes

13. What About Latency?
Continuous processing mode to run without
microbatches
• <1 ms latency (same as per-record streaming systems)
• No changes to user code
• Proposal in SPARK-20928
Key idea: same API can target both streaming & batch
Find out more in today’s deep dive

14. Expanding Spark to New
Areas
Structured Streaming
Deep Learning
1
2

15. Deep Learning has Huge
PotentialUnprecedented ability to work with unstructured
data such as images and text

16. But Deep Learning is Hard to
UseCurrent APIs (TensorFlow, Keras, etc) are low-level
• Build a computation graph from scratch
Scale-out requires manual parallelization
Hard to use models in larger applications
Very similar to early big data APIs

17. Deep Learning on Spark
Image support in MLlib: SPARK-21866 (Spark 2.3)
DL framework integrations: TensorFlowOnSpark,
MMLSpark, Intel BigDL
Higher-level APIs: Deep Learning Pipelines

18. New in TensorFlowOnSpark
Library to run distributed TF on Spark clusters & data
• Built at Yahoo!, where it powers photos, videos & more
Yahoo! and Databricks collaborated to add:
• ML pipeline APIs
• Support for non-YARN and AWS clusters
github.com/yahoo/TensorFlowOnSpark
talk
tomorrow
at 17:00

19. Deep Learning Pipelines
Low-level DL frameworks are powerful, but common
use cases should be much simpler to build
Goal: Enable an order of magnitude more
users to build production apps using deep
learning

20. Deep Learning Pipelines
Key idea: High-level API built on ML Pipelines model
• Common use cases are just a few lines of code
• All operators automatically scale over Spark
• Expose models in batch, streaming & SQL apps
Uses existing DL engines (TensorFlow, Keras, etc)

21. Example: Using Existing
Modelpredictor = DeepImagePredictor(inputCol="image",
outputCol="labels",
modelName="InceptionV3")
predictions_df = predictor.transform(image_df)
SELECT image, my_predictor(image) AS labels
FROM uploaded_images

22. Example: Model Search
est = KerasImageFileEstimator()
grid = ParamGridBuilder()
.addGrid(est.modelFile, ["InceptionV3", "ResNet50"])
.addGrid(est.kerasParams, [{'batch': 32}, {'batch': 64}])
.build()
CrossValidator(est, eval, grid).fit(image_df)
InceptionV3
batch size 32
ResNet50
batch size 32
InceptionV3
batch size 64
ResNet50
batch size 64
Spark
Driver

23. Deep Learning Pipelines
DemoSue Ann Hong